ChatGPT is OpenAI’s AI assistant for conversations, research, writing, coding, and other everyday tasks. Its token and usage limits vary by plan, model, and response mode. Free and Go accounts are moving to GPT-5.6 Luna as the default model, while Plus and Pro accounts are receiving GPT-5.6 Sol for both Instant replies and deeper reasoning.
Plus and Pro include a reasoning slider that controls how much thought GPT-5.6 Sol applies to each response. Existing ChatGPT context limits depend on the response mode: 32K for Instant and up to 256K for manually selected reasoning on Plus, and 128K for Instant and up to 400K for manually selected reasoning on Pro. API requests follow model-specific limits.
OpenAI API limits are model-specific. GPT-5.6 Sol, Terra, and Luna each support a 1,050,000-token context window and 128,000 output tokens. GPT-4o supports a 128,000-token context window and 16,384 output tokens. GPT-4.1 supports 1,047,576 context tokens and 32,768 output tokens.
ChatGPT and the OpenAI API apply several types of limits. The context window controls the total material available to one request, and maximum output caps one response. ChatGPT message allowances and file-upload quotas control product usage. API rate limits control throughput over time.
ChatGPT Free, Go, Plus, and Pro Token Limits
ChatGPT is rolling out GPT-5.6 Luna as the default model for Free and Go accounts. Plus and Pro receive GPT-5.6 Sol across Instant responses and manually selected reasoning. The applicable token limit depends on both the subscription plan and the selected response mode.
| ChatGPT plan | Default Model | Manual reasoning | Context limits |
|---|---|---|---|
| Free | GPT-5.6 Luna | Think button | – |
| Go | GPT-5.6 Luna | Think button | – |
| Plus | GPT-5.6 Sol | Medium and High | 32K for Instant; up to 256K for manually selected reasoning, divided into 128K input and 128K maximum output |
| Pro | GPT-5.6 Sol | Medium, High, Extra High, and GPT-5.6 Sol Pro | 128K for Instant; up to 400K for manually selected reasoning, divided into 272K input and 128K maximum output |
Medium and High use GPT-5.6 Sol on Plus. Pro adds Extra High with GPT-5.6 Sol and the Pro option powered by GPT-5.6 Sol Pro. The reasoning slider is available on web, mobile, and desktop. Automatic switching from Instant to Medium does not consume the allowance for manually selected reasoning.
GPT-5.6 Luna powers the default Free and Go chat experience during the rollout and has no manual selector in standard conversations. Standard Chat also has no GPT-5.6 Terra selector. This Chat-focused Sol update applies only to Chat; Work and Codex keep their existing model access. The OpenAI API provides Sol, Terra, and Luna. ChatGPT subscription charges and API billing are separate.
OpenAI API Token Limits by Model
API token limits belong to the exact model ID used in the request. The context window limits the total request size. The maximum output value limits generated tokens. API usage tiers can raise throughput limits without changing either model specification.
| OpenAI API model | Context window | Max output tokens | Current API status |
|---|---|---|---|
| GPT-5.6 Sol | 1,050,000 | 128,000 | Current model; the gpt-5.6 alias routes to Sol |
| GPT-5.6 Terra | 1,050,000 | 128,000 | Current model |
| GPT-5.6 Luna | 1,050,000 | 128,000 | Current model |
| GPT-4.1 | 1,047,576 | 32,768 | Available API model |
| GPT-4o | 128,000 | 16,384 | Available API model |
| GPT-4o mini | 128,000 | 16,384 | Available API model |
GPT-5.6 Sol, Terra, and Luna share the same context window and maximum output. Capability and API price distinguish the three models.
Current prices are listed in the AI LLM API pricing reference.
GPT-4o Max Output Tokens and Context Window
The gpt-4o API model has a 128,000-token context window and a 16,384-token maximum output. The 128K figure describes the total request capacity. It does not describe the length of the generated answer.
A request containing 120,000 input tokens leaves about 8,000 tokens inside the 128,000-token context window for the response. The 16,384-token maximum output does not override that shared context limit. In this example, the remaining context capacity controls the actual response length.
Context Window vs Maximum Output Tokens
A context window is the total token capacity available to one model request. It includes instructions, user input, supplied conversation history, retrieved passages, tool results, and the generated answer. Reasoning models also use part of this capacity for internal reasoning tokens.
Maximum output tokens apply only to generated tokens. A model can have a one-million-token context window and a much smaller output limit. GPT-4.1, for example, accepts up to 1,047,576 tokens in its context and generates up to 32,768 output tokens.
Input must leave enough capacity for reasoning and the requested answer. Filling nearly the entire context window with source material can reduce the available output or cause the request to exceed the model limit.
The maximum output value is a ceiling. A response can end earlier after completing the requested format, reaching a stop condition, triggering a safety rule, or using the configured output budget for reasoning and other generated tokens.
ChatGPT Message Limits Are Different from Token Limits
A message allowance controls how often an account can use a model during a stated period. A context window controls the amount of information available inside one conversation request. Reaching a message allowance can temporarily change model access even when the current conversation uses little of its available context.
The staged Free and Go rollout moves text chats to GPT-5.6 Luna with no message rate limit, subject to abuse guardrails. The Think button also uses GPT-5.6 Luna for questions that need more reasoning. File uploads, images, and other ChatGPT tools keep their own limits.
Manually selected Medium, High, and Extra High modes use GPT-5.6 Sol on eligible plans. These modes keep the existing plan-specific reasoning allowances. When a GPT-5.6 reasoning allowance is exhausted, ChatGPT can continue with GPT-5.4 Thinking mini when that fallback is available.
ChatGPT File Upload Limits
File-upload limits govern file size, document tokens, upload frequency, and stored files. These limits do not increase the context window used for a ChatGPT response.
| File limit | Current value | Scope |
|---|---|---|
| Maximum file size | 512 MB per file | Files uploaded to a GPT or ChatGPT conversation |
| Text and document token cap | 2 million tokens per file | Text and document files; spreadsheets use a different size rule |
| Spreadsheet size | Approximately 50 MB | Depends on the size of each row |
| Image size | 20 MB per image | Uploaded images |
| Rolling upload allowance | Up to 80 files every 3 hours | The allowance can be reduced during peak periods |
| Free-plan upload allowance | 3 file uploads per day | ChatGPT Free |
| Stored files | 25 GB per user; 100 GB per organization | Shared across supported ChatGPT file surfaces |
Token Limits vs OpenAI API Rate Limits
Token limits control the size of one API request and its generated response. Rate limits control the amount of API traffic processed during a period. Common rate-limit measurements include requests per minute, tokens per minute, requests per day, and batch queue capacity.
A higher API usage tier can increase throughput. The selected model uses its published context window and output limit at every tier. An application can therefore have unused tokens-per-minute capacity and reject one request that exceeds the model context.
See OpenAI API rate limits for current RPM, TPM, daily, and batch-limit details.
What Counts as a Token
Tokens are the units processed by language models. A token can represent a complete word, part of a word, punctuation, whitespace, a number, or a piece of code. English text averages about four characters or three-quarters of a word per token.
Code, JSON, URLs, tables, logs, and multilingual text can use tokens at different rates. Two documents with the same character count can therefore produce different token counts.
Use a model-compatible tokenizer when a request approaches the model limit. Character estimates are useful for planning, yet they cannot confirm that a large code sample, structured dataset, or multilingual document fits.
Reasoning Tokens and the API Output Budget
Reasoning models spend tokens on internal reasoning before producing visible text. These tokens reduce the generated-token budget available to the visible response. Complex coding, mathematics, planning, and research tasks can therefore consume a substantial output budget before the final answer appears.
In the Responses API, max_output_tokens caps the combined total of reasoning tokens and visible output tokens. When the model uses the full configured budget before completing the answer, the response returns an incomplete status with max_output_tokens as the reason.
Reserve output capacity for reasoning models before sending a large prompt. Inspect the response usage object to see reasoning and output token counts, then adjust the prompt size or configured output budget for later requests.
How to Work Within Token Limits
- Remove duplicate instructions and unrelated examples.
- Split long documents into named sections or page ranges.
- Retrieve only the passages required for the current question.
- Replace long conversation history with a concise project brief.
- Reserve enough context for reasoning and the final answer.
- Generate long reports, code, or structured data in planned sections.
- Use a model whose context and output limits match the workload.
- Pin a dated API snapshot to keep production behavior stable.
Older and Deprecated OpenAI Model Limits
| Model | Context window | Max output tokens | Status |
|---|---|---|---|
| GPT-4 Turbo | 128,000 | 4,096 | Deprecated |
| GPT-4 | 8,192 | 8,192 | Deprecated |
| GPT-3.5 Turbo | 16,385 | 4,096 | Deprecated legacy model |
Real-world Token-Limit Examples
Analyzing a Long PDF in ChatGPT
A PDF that fits below the two-million-token file cap can contain more text than the selected ChatGPT mode can use in one response. Ask for analysis of specific chapters or page ranges. Save concise section summaries and use those summaries for the final comparison or report.
Generating a Long API Report
If the required report exceeds the selected model’s maximum output, divide it into named sections and generate one section per request. Store each completed section in the application, then run a final editing pass over an outline or condensed draft to check terminology, structure, and cross-section consistency.
Working with a Large Codebase
Start with the task, relevant source files, dependency definitions, exact error output, and a short project map. Generated files, build artifacts, vendored packages, and full logs consume context without helping most first-pass coding tasks.
Add the next file or log range after the model identifies a concrete dependency. This keeps enough context for analysis and code changes.
See Best AI Coding Agents for tools designed around repository workflows.
FAQs
Can I set max_output_tokens above the model’s maximum output?
No. The request must stay within the selected model’s output cap. A larger parameter value does not extend the model specification.
Does starting a new ChatGPT conversation reset a message allowance?
No. Message allowances reset according to the plan’s time window. Opening a new conversation does not restore access to a model whose allowance has been used.
Does asking ChatGPT or the API to continue bypass the maximum output limit?
Each response receives its own output budget. A continuation can produce the next part, though the retained conversation and previous output also consume context. Long results work better when the prompt defines section boundaries in advance.
Should an API application use a model alias or a dated snapshot?
An alias receives model updates automatically. A dated snapshot provides a stable model version for applications that require consistent behavior and verified limits.
Related Resources
- OpenAI API Rate Limits: Requests, Tokens, Tiers, and Batch Limits
- AI LLM API Pricing: OpenAI, Gemini, Claude, and More
- OpenAI and ChatGPT Timeline: Model Releases and Product Updates
- OpenAI API Model Catalog
- OpenAI Token Counting Guide
Changelog
August 6, 2026: GPT-5.6 Luna began replacing GPT-5.5 Instant as the Free and Go default. The staged rollout adds unlimited text chats and a Think button for those plans. Plus and Pro receive GPT-5.6 Sol across Instant and deeper reasoning with a reasoning slider.
August 4, 2026: Updated ChatGPT Free, Plus, and Pro model access and context limits; added GPT-5.6 reasoning availability; refreshed GPT-4o output limits, ChatGPT file-upload limits, reasoning-token behavior, and deprecated model status.
July 21, 2026: Updated current API limits for GPT-5.6 Sol, Terra, and Luna; corrected GPT-4o, GPT-4o mini, and GPT-4.1 token-limit coverage.
In this article, the “Token Limit in GPT-4” heading, and the “Max tokens” in the table are incorrect. The numbers do not indicate the Max tokens. They are Context Window.
Fixed. I forget to add the ‘Max Output Tokens’ column. Thanks for your feedback.